What the documents look like
Brand member operations in the private domain works from four document families: member tag tables, outreach plans, coupon rules, and privacy authorization records. These come from the CRM system, enterprise messaging, POS transaction records, and mini-program behavior data. In a cross-border and multilingual setting, the same member tag table or coupon rule may exist in both a Chinese and an English version. Semantic retrieval can match across languages, so a Chinese question can reach an English knowledge point.
Two or three facts shape this page. First, member tag tables and coupon rules carry identifiers such as member numbers, coupon codes, and authorization numbers, and those identifiers are the kind of content that exact full-text retrieval handles well. Second, privacy authorization records hold status values and date ranges, so a retrieval hit only shows that a record exists; it does not show that an authorization was executed. Third, the same member tag or coupon rule may have parallel language versions, which makes cross-language semantic matching useful but also makes it necessary to check which version a retrieved passage came from. This page covers the retrieval and recall step only. It does not infer batch synchronization or transaction-triggered mechanisms.
Constraints at the retrieval step
The knowledge base search node can link one or more knowledge bases that use the same vector model. Knowledge bases built on different vector models cannot be linked together; the model identifier has to be checked first and non-matching bases excluded. The node outputs cited content as an array, and the length can be 0. Even when nothing is found, the output path still runs through.
Semantic retrieval, full-text retrieval, and hybrid retrieval suit different material. Coupon codes, authorization numbers, and other exact queries suit full-text retrieval. Natural-language questions suit semantic retrieval. Hybrid retrieval uses both and merges results through RRF. The minimum relevance value takes effect only under semantic retrieval or when result reranking is used; it is a filter threshold and does not perform permission or entity isolation. Retrieval itself does not constitute authorization. Image retrieval needs image input and matching model capability, and reranking needs a model and must be enabled in settings. The paths drawn in the overall retrieval diagram do not mean every request executes all of them.
Hypothetical drill sample. The source text to be checked is: "Privacy authorization record for member number M-2041: status signed, signing date 12 March 2024, authorization scope cross-border marketing outreach. Outreach plan for member number M-2041: status pending review, submission date 15 March 2024." The draft sentence to be judged is: "M-2041 has completed cross-border outreach authorization and entered the execution stage." This draft sentence merges two different fields, signed and pending review, into one conclusion, and pending review does not constitute execution.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this drill | Reason and comparison object |
|---|---|---|
| Linked knowledge bases | Link only member material bases that use the same vector model | Check each base's vector model identifier first and exclude bases with a different model |
| Search mode | Run semantic retrieval and full-text retrieval separately on the same question | Compare how the two modes hit an authorization number against a natural-language question |
| Citation limit | Start at the token volume needed to cover one complete authorization record | Test with material of different lengths to see whether content stays complete, then adjust |
| Minimum relevance | Start at 0.5, effective only under semantic retrieval or reranking | Label relevant and irrelevant samples first, then compare result changes after filtering |
| Question optimization | Enable for multi-turn follow-ups, disable for exact identifier questions | Compare the original question with the optimized question and check whether the subject was altered |
| Result reranking | Enable when candidates are numerous, according to settings | Compare rerank scores against vector scores for the same candidate |
Three common errors
- Treating "pending review" as "executed." In the sample, the outreach plan status is pending review, while the draft sentence says the member entered the execution stage. This merges values from different fields into one conclusion. Check: input the draft sentence and the source text, compare the status field values, and observe whether pending review is written as execution.
- Treating a retrieval hit as an established authorization. Finding a privacy authorization record only shows that the material exists; it does not constitute authorization execution. Check: input one retrieval result, compare it with the authorization status field, and observe whether the recalled content is written directly as a completed authorization.
- Linking knowledge bases across vector models. The sample requires the same vector model. If bases with different models are linked together, the only valid step is to check identifiers and exclude non-matching bases. Check: input the model identifiers of two bases, compare whether they match, and observe whether a non-matching base is excluded.
How to verify the setup
- For the sample draft sentence, input "M-2041 has completed cross-border outreach authorization and entered the execution stage," compare it with the source text where authorization status is signed and outreach plan status is pending review, and observe whether the draft sentence is judged unsupported, because pending review does not equal execution.
- For empty-result handling, input a paper retrieval result of [], compare whether the downstream node continues on an array length of 0, and observe whether the output path is not interrupted. Do not claim that a fictional rule retrieval necessarily returns [].
- For cross-model linking, input the vector model identifiers of two knowledge bases, compare whether they are the same, observe whether the base with a different identifier is excluded from joint linking, and record the actual observation without inferring runtime results.